Neural Representation of Multiple Languages in Polyglots: fMRI Evidence for a Shared Language Network

This supplementary report investigates how polyglots’ language networks respond to native (L1), familiar non-native (L2–L4), unfamiliar related (URLs), and unfamiliar unrelated (UULs) languages compared to a perceptually matched quilted-speech control. It examines robustness of findings across materials (Bible vs. Alice), participant age, scanning runs, hemispheric dominance, and participant subgroups (e.g., excluding early balanced bilinguals, left-handed individuals, and cases with stimulus-selection errors). It also compares native-language responses between polyglots and non-polyglot controls and characterizes right-hemisphere homotopic responses and Multiple Demand (MD) network responses. An objective lexical distance metric (Beaufils & Tomin, 2020) validates the selection of unfamiliar related vs. unrelated languages.

- Participants: 34 polyglots and 86 non-polyglot controls (for native-language comparisons). Age range 19–71 years (M=35.0, SD=13.9). Handedness varied; two left-handed participants. One participant (UID 511) showed right-hemisphere language lateralization (LI = -0.854). - Experimental conditions: L1 (native), familiar non-native L2–L4 (L2 highest non-native proficiency; L3/L4 lower), two unfamiliar related languages (URLs) selected to be related to a participant’s high-proficiency language(s), two unfamiliar unrelated languages (UULs), and a quilted-speech control condition. - Materials: Two sets used across participants: publicly available Bible audio stories (n=18) and passages from Alice in Wonderland (n=16). - fMRI and ROI definition: Language functional ROIs (fROIs) defined via Sentences > Nonwords contrast in an English localizer. In additional analyses, L1 > Quilted contrast was also used to define the language network, yielding similar results. For visualization of overlap maps, the top 10% of voxels in left-hemisphere parcels most responsive to Language > Quilts were displayed. Right-hemisphere homotopic language areas were also examined. - Analysis models: • Robustness to materials: LME EffectSize ~ Materials + (1|Condition) + (1|Participant) + (1|ROI). • Robustness to age: LME EffectSize ~ Age + (1|Condition) + (1|Participant) + (1|ROI). • Condition effects within each language fROI: LME EffectSize ~ Condition + (1|Participant) fitted per ROI. • Between-group (polyglots vs. non-polyglots) native-language comparison performed for each fROI (uncorrected and Bonferroni-corrected p-values reported). • Robustness checks across scanning runs, participants (split-half correlations across random halves, 1,000 iterations), materials, and participant exclusions (e.g., excluding early balanced bilinguals; excluding participants with L1 proficiency <20; excluding left-handers; and using language-dominant hemisphere). - Inter-language distance validation: Using Beaufils & Tomin (2020) lexical-distance metric (scale 1–100), distances were computed between URLs and their related languages and, for UULs, averaged distances to L1–L4 per participant. Independent-samples t-test compared URLs vs. UULs distances. - Data availability and IDs: Participant UIDs link to OSF (https://osf.io/3he75/). Tables/Figures S1–S15 detail demographics, lateralization indices, language sets, ROI-wise statistics, and subgroup/robustness analyses. - Noted stimulus-selection deviations: Two participants had errors in related-language selection (UIDs 383, 867; affected data excluded as specified). One participant (UID 1041) had only one URL and a modified set of familiar languages due to participation in a different study; extra familiar-language conditions were excluded to maintain design consistency.

- Validation of unfamiliar language selection by distance metric: URLs (n=64): M=23.25, SD=7.95; UULs (n=67): M=89.9, SD=2.99; independent-samples t(79.8) = -62.5, p<0.001, confirming URLs are lexically close and UULs are distant. - Robustness to materials: LME Materials effect not reliable (β = -0.137, p = 0.603). The overall condition pattern is robust across Bible vs. Alice materials. - Robustness to age: LME Age effect not reliable (β = -0.005, p = 0.617); condition pattern does not vary with age. - Condition responses vs. quilts control in language fROIs (Table S4): After Bonferroni correction across five fROIs, significant responses were observed for L1–L4 and L6 in all fROIs; L5 significant for all fROIs except LH IFGorb; L7/L8 significant in some fROIs. Overall, all language conditions elicited reliable responses above the quilt control (ps < 0.001 reported broadly). - Right hemisphere homotopic language network: All language conditions showed significant responses above quilts (ps < 0.001). However, the familiar vs. unfamiliar difference was less pronounced in RH than LH; LH mean responses: familiar 1.81 vs. unfamiliar 0.95; RH: familiar 1.45 vs. unfamiliar 1.04; interaction p = 0.010. - Native-language response: Polyglots showed a numerically lower native-language response than 86 non-polyglot controls (1.79% vs. 2.46% BOLD signal change; β = -0.53, p = 0.063). In ROI-wise comparisons (Table S5), all fROIs except LH PostTemp survived Bonferroni correction for the polyglot vs. non-polyglot native-language difference. - Multiple Demand (MD) network: Overall responses to auditory language are low. Familiar, related, unrelated languages, and quilts are above baseline (weak), but native language elicits, on average, a below-baseline response; the Related condition elicits the highest MD response among language conditions. - Robustness across runs and participants: High reproducibility; split-half correlations across random halves (1,000 iterations) show strong consistency (mean correlation M = 0.89; SD = 0.65; range 0.57–0.99). Patterns are robust when excluding early balanced bilinguals (L2 also native), participants with L1 proficiency <20, left-handers, and when analyzing language-dominant hemisphere. - Spatial activation extent: At a fixed threshold (p < 0.001 uncorrected, L1 > Quilts), polyglots show less spatially extensive activation than non-polyglots for native-language processing, mirroring lower magnitude responses in individually defined fROIs.

The supplementary analyses converge on a robust hierarchy of responses in the polyglot language network: native and familiar languages elicit the strongest activity, unfamiliar related and unrelated languages elicit weaker activity, and all language conditions exceed the quilts control. This hierarchy holds across materials, age, runs, and various participant exclusions, indicating that it reflects stable properties of the polyglot language system rather than task- or sample-specific artifacts. The right-hemisphere homotopic language areas show qualitatively similar but attenuated differentiation between familiar and unfamiliar languages, highlighting LH dominance for fine-grained language proficiency effects. Relative to non-polyglots, polyglots tend to exhibit lower native-language responses in the language network, a pattern present across most fROIs and accompanied by less extensive activations at a fixed threshold. The MD network shows generally low engagement for auditory language, with a notable suppression for native-language processing and greater responses for non-native/related language conditions, consistent with higher control/effort demands outside the core language system. The lexical distance metric analysis confirms that URLs are indeed closely related to familiar languages whereas UULs are distant, validating the experimental manipulation of language relatedness.

This supplementary report demonstrates that polyglots’ language networks exhibit robust and graded responses across native, familiar, and unfamiliar (related and unrelated) languages, consistently above a quilted-speech control, and that these patterns are stable across materials, age, runs, and analytic choices. Polyglots show a trend toward reduced native-language responses compared to non-polyglots across most language fROIs, with less spatially extensive activation at a fixed threshold. Right-hemisphere homotopic areas show similar, though weaker, sensitivity to language familiarity, and the MD network displays low responses overall with native-language suppression. The inter-language distance analysis objectively validates the selection of unfamiliar related vs. unrelated languages. Potential future directions were not detailed in the provided text.

- Stimulus selection issues for a subset of participants: errors in related-language selection (UIDs 383 and 867); affected data were excluded as specified. UID 1041 had only one URL and a modified familiar-language set due to participation in a different study; extra familiar conditions were excluded to maintain design consistency. - Materials availability constrained the choice of L3/L4 and related languages in some cases. - Sample characteristics: small polyglot sample (n=34) and only two left-handed participants; one participant had right-hemisphere language dominance (analyses accounted for this by using RH fROIs as appropriate). - Reported p-values in tables are uncorrected unless noted; multiple-comparisons control highlighted for specific cases (Bonferroni across five fROIs).